| Radio signal identification plays an important role in the fields of electronic reconnaissance and civil radio monitoring.At present,the most popular smart radio signal recognition methods are based on deep learning algorithms,such as convolutional neural networks.Typically,traditional convolutional neural networks require large labeled datasets.However,in the fields of electronic reconnaissance and civil radio monitoring,there are always some radio signals that interfere with the radio environment and transmit transiently,such as "black radio stations".Such radio signals are not easy to collect,that is,the labeled signal samples are quite rare,which makes the deep learning model unable to effectively learn the high-dimensional features of the signal,so that these radio signals cannot be effectively identified.Therefore,it is necessary to study the few-shot radio signal inference recognition technology to improve the accuracy of radio signal reasoning and recognition in the case of small samples.The innovations of this paper are as follows:(1)Traditional radio signal inference recognition algorithms based on convolutional neural networks require a large number of labeled signal samples to perform a good fit on signal dataset.However,when the number of labeled signal samples is small,the network cannot fit the signal data well,resulting in a low signal inference recognition accuracy of the network model.In view of the above problems,and combined with the characteristics of transfer learning,which can transfer the signal data knowledge in other fields(source domain)to the target domain signal data set,thereby improving the efficiency of the model learning target domain signal data set,this paper innovatively proposes an attention transfer learning algorithm.That is,the basic CNN(convolutional neural network)network is improved based on the improved convolutional block attention mechanism(CBAM)and transfer learning technology,and the algorithm is tested under the condition of no signal fading and signal fading.The results show that in the case of no signal fading,when the number of samples of each type of signal is 50,the signal inference recognition accuracy of the model which corresponding to the attention transfer learning algorithm is improved by about 19%compared with the basic CNN network;And in the case of signal fading,when the number of samples of each type of signal is 50,the signal inference recognition accuracy of the model which corresponding to the attention transfer learning algorithm is about 18%higher than that of the basic CNN network.(2)Traditional transfer learning algorithms need to learn from signal datasets in other fields to effectively improve the signal inference recognition accuracy of the model under small samples.However,when there is little or no signal data in other fields,the accuracy of the signal recognition algorithm based on transfer learning is low.Aiming at this problem,combined with the characteristics of Siamese network,which can transform the multi-classification problem into a binary classification problem,thereby reducing the complexity of the network model,this paper innovatively proposes a deep convolutional block attention Siamese network.That is,improve the basic CNN network based on the improved CBAM attention fusion Siamese network,and experiments on this network based on the no-fading signal dataset.The results show that when the number of signal samples per class is 40,the signal inference recognition accuracy of the deep convolutional block attention Siamese network is improved by about 29%compared with the basic CNN network. |
